Re: [PATCH v6 09/10] hisi_acc_vfio_pci: Add support for VFIO live migration

[Alex Williamson]

On Tue, 1 Mar 2022 16:39:38 -0400
Jason Gunthorpe <jgg@xxxxxxxxxx> wrote:

> On Tue, Mar 01, 2022 at 12:30:47PM -0700, Alex Williamson wrote:
> > Wouldn't it make more sense if initial-bytes started at QM_MATCH_SIZE
> > and dirty-bytes was always sizeof(vf_data) - QM_MATCH_SIZE?  ie. QEMU
> > would know that it has sizeof(vf_data) - QM_MATCH_SIZE remaining even
> > while it's getting ENOMSG after reading QM_MATCH_SIZE bytes of data.  
> 
> The purpose of this ioctl is to help userspace guess when moving on to
> STOP_COPY is a good idea ie when the device has done almost all the
> work it is going to be able to do in PRE_COPY. ENOMSG is a similar
> indicator.
> 
> I expect all devices to have some additional STOP_COPY trailer_data in
> addition to their PRE_COPY initial_data and dirty_data
> 
> There is a choice to make if we report the trailer_data during
> PRE_COPY or not. As this is all estimates, it doesn't matter unless
> the trailer_data is very big.
> 
> Having all devices trend toward a 0 dirty_bytes to say they are are
> done all the pre-copy they can do makes sense from an API
> perspective. If one device trends toward 10MB due to a big
> trailer_data and one trends toward 0 bytes, how will qemu consistently
> decide when best to trigger STOP_COPY? It makes the API less useful.
>
> So, I would not include trailer_data in the dirty_bytes.

That assumes that it's possible to keep up with the device dirty rate.
It seems like a better approach for userspace would be to look at how
dirty_bytes is trending.  A zero value and a steady state value are
equivalent, there's nothing more to be gained by further iterations.  If
the value is trending down, it might be worthwhile to iterate in
PRE_COPY a while longer.  If the value is trending up, it might be time
to cut to STOP_COPY or abort the migration.

If we exclude STOP_COPY trailing data from the VFIO_DEVICE_MIG_PRECOPY
ioctl, it seems even more of a disconnect that when we enter the
STOP_COPY state, suddenly we start getting new data out of a PRECOPY
ioctl.

BTW, "VFIO_DEVICE" should be reserved for ioctls and data structures
relative to the device FD, appending it with _MIG is too subtle for me.
This is also a GET operation for INFO, so I'd think for consistency
with the existing vfio uAPI we'd name this something like
VFIO_MIG_GET_PRECOPY_INFO where the structure might be named
vfio_precopy_info.

> Estimating when to move on to STOP_COPY and trying to enforce a SLA on
> STOP_COPY are different tasks and will probably end up with different
> interfaces.
> 
> I still think the right way to approach the SLA is to inform the
> driver what the permitted time and data size target is for STOP_COPY
> and the driver can proceed or not based on its own internal
> calculation.

So if we don't think this is the right approach for STOP_COPY, then why
are we pushing that it has any purpose outside of PRECOPY or might be
implemented by a non-PRECOPY driver for use in STOP_COPY?
 
> > useful yet and you don't want to add dead kernel code, then let's
> > define that this ioctl is only available in the PRE_COPY* states and
> > returns -errno in the STOP_COPY state.  
> 
> I'm OK with that, in acc it is done by checking migf->total_bytes >
> QM_MATCH_SIZE during the read fop
> 
> > devices in STOP_COPY and let's also define if there's actually anything
> > userspace can infer about remaining STOP_COPY data size while in
> > PRE_COPY* via this ioctl.  For example, is dirty-bytes zero or the
> > remaining data structure size?  
> 
> If we keep it then I would say it doesn't matter, userspace has to sum
> the two values to get the total remaining length estimate, it is just
> a bit quirky.

For the reasons above, I just can't figure out why wouldn't decide that
use of this outside of PRECOPY is too quirky to bother with.  Thanks,

Alex